考虑非邻近节点空间相关性的交通流预测模型  

Traffic prediction model considering spatial correlation of non-neighboring nodes

作　　者：闫光辉[1] 李鸿涛 张斌[1] 常文文 Yan Guanghui;Li Hongtao;Zhang Bin;Chang Wenwen(School of Electronic&Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)

机构地区：[1]兰州交通大学电子与信息工程学院,兰州730070

出　　处：《计算机应用研究》2025年第3期825-833,共9页Application Research of Computers

基　　金：国家自然科学基金资助项目(62062049);中央引导地方科技发展资金资助项目(22ZY1QA005);甘肃省教育厅青年博士项目(2023QB-038);2024年研究生教育教学质量提升工程建设项目(JG202418)。

摘　　要：针对现有的交通流预测模型存在难以对非邻近节点之间的时空相关性显式建模的问题,提出一种新的利用超图表征空间相关性的超图卷积神经网络模型(double attention hypergraph convolution neural network,A2HGCN)。首先,通过寻找节点之间的相似关系构造超边,利用节点之间的连接关系构造超图;然后提出一个超图卷积模型,其中利用超图卷积和将超图线扩展为图后利用线图卷积来捕获潜在的空间相关性;再利用融合双层注意力机制的卷积长短时记忆网络捕获时间相关性,最后得出预测结果。在数据集PEMS-BAY中,A2HGCN方法的评价指标MAE、MAPE和RMSE在预测步长为15 min时为1.223、2.617%、2.547,30 min时为1.554、3.541%、3.420,60 min时为1.867、4.578%、4.224。在数据集PEMSM中,该方法的评价指标MAE、MAPE和RMSE在预测步长为15 min时为1.858、4.385%、3.339,30 min时为2.374、5.775%、4.362,60 min时为3.046、7.713%、5.479。结果表明,该方法在不同预测步长下均优于基线模型,验证了考虑非邻近节点之间的时空相关性对于提高交通预测准确性的有效性。由此可得,超图卷积神经网络在捕获时空相关性方面具有优势。This paper proposed a novel double attention hypergraph convolution neural network(A2HGCN)to address the challenge of explicitly modeling spatiotemporal correlations between non-adjacent nodes in existing traffic flow prediction mo-dels.The method used the construction of hyperedges based on similarities between nodes and created hypergraphs through node connectivity to represent spatial correlations.It proposed a hypergraph convolution model,which employed hypergraph convolution and line graph convolution after expanding hypergraph lines into graphs to capture potential spatial correlations.Meanwhile,it used a convolutional long short-term memory network integrated with a double attention mechanism to capture temporal features.The algorithm made predictions based on spatial and temporal features.In the PEMS-BAY dataset,the A2HGCN method achieved evaluation metrics of MAE,MAPE,and RMSE at a prediction step of 15 minutes as 1.223,2.617%,and 2.547,respectively,at 30 minutes as 1.554,3.541%,and 3.420,and at 60 minutes as 1.867,4.578%,and 4.224.In the PEMSM dataset,the method achieved evaluation metrics of MAE,MAPE,and RMSE at a prediction step of 15 minutes as 1.858,4.385%,and 3.339,at 30 minutes as 2.374,5.775%,and 4.362,and at 60 minutes as 3.046,7.713%,and 5.479.The results demonstrate that the proposed method outperforms baseline models at different prediction steps,validating the effectiveness of considering spatiotemporal correlations between non-adjacent nodes in enhancing traffic prediction accuracy.It is concluded that hypergraph convolutional neural networks have an advantage in capturing spatiotemporal correlations.

关 键 词：交通流预测 超图理论 图卷积网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]